I work on getting large models to run fast and cheap: serving, quantization, and the environments used to train them with RL. I like digging into how these systems actually work, which usually means building on vLLM and Hugging Face projects and contributing back what I learn.
What I've been working on
- vLLM llm-compressor: KV-cache scale calibration fix
- vLLM humming: f16 GEMM epilogue bias fix, plus a shared-memory overflow fix on SM86
- Hugging Face OpenEnv: closed a reward-hacking hole in five sandbox envs
- Hugging Face Repo2RLEnv: stopped the oracle patch leaking into agent images
- Hugging Face funes: 1.8× faster indexing, 6× less memory



